Microsoft is targeting a 40% increase in tokens-per-megawatt efficiency from a custom AI accelerator by 2027, according to a report carried by The Cryptonomist.

The headline places energy efficiency — measured as generated tokens divided by electricity consumed — at the centre of Microsoft's AI silicon roadmap. A token is the basic unit of text a model reads or writes. Tokens per megawatt quantifies how much inference work a chip performs per megawatt-hour drawn.

Microsoft's stated target translates the abstract goal of "more efficient AI" into a single comparable number. The 40% improvement measures against a stated baseline, although the report does not specify which chip serves as the reference.

For a fleet operator, a 40% efficiency gain compounds quickly. A 100-megawatt cluster producing X tokens per hour today would, under the target, produce 1.4 × X tokens per hour at the same draw. Equivalently, the same workload would run on roughly 29% less electricity — 1 ÷ 1.4 ≈ 0.71.

The metric itself is rare in chip marketing. Vendors typically publish teraflops, transistor counts, memory bandwidth, or inference latency in milliseconds. Tokens per megawatt fuses two numbers — output quantity and energy — into a single ratio that captures both throughput and operating cost.

The 2027 horizon gives Microsoft until that year to deliver the gains described. The Cryptonomist report did not include direct quotes from Microsoft executives and did not confirm whether the 40% figure originated with the company or with third-party analysts.

Microsoft has not, in available reporting, disclosed the chip's process node, memory configuration, foundry partner, or intended deployment scale. The published number — 40% more tokens per megawatt by 2027 — is a target, not a measured result, and the comparison baseline remains unspecified.

What does "tokens per megawatt" actually measure?

The metric divides text units a model processes or generates by the electricity consumed, typically per hour. A higher number means more useful output per unit of energy. It bundles raw computational throughput and steady-state power draw into a single operational ratio that data-centre operators can compare across vendors.

Why benchmark AI on energy rather than flops?

FLOPS quantify theoretical compute capacity but ignore memory access, interconnect overhead, and idle power. For inference workloads running continuously on large clusters, the binding constraint is usually available megawatts rather than peak silicon performance. Tokens per megawatt captures the speed-and-consumption combination that determines real-world operating cost at hyperscale.

What remains undisclosed?

• The reference chip used for the 40% comparison
• Fabrication process and memory architecture
• Whether the figure targets training, inference, or both
• Intended deployment region or partner
• Whether Microsoft engineering set the target or external analysts projected it

The published report leaves the 40% claim as an unverified ceiling. Until Microsoft — or any foundry partner — confirms the baseline chip, the workload, and the measurement methodology, the target remains a benchmark ambition rather than a tested result.